Abstract—This paper proposes a method for pedestrian object

detection by using Centroid Neural Network (CNN). SIFT(Scale

Invariant Feature Transform) is used to produce keypoint feature

extracted from image data and the keypoints are used to discriminate a

scene with pedestrian objects from a scene without pedestrian objects.

Experiments on INRIA Person dataset show that the keypoint features

extracted by using SIFT are useful for pedestrian object detection

problems and the proposed CNN classifier can detect pedestran object

effectively.

Keywords—pedestrian object, neural network, feature, image

data.

I. INTRODUCTION

YPICALLY, there are two different procedures for various

pattern recognition problems including pedestrian objects:

feature extraction procedure and classifier design procedure.

For feature extraction procedure for pedestrian object detection,

there are two different categories of approaches in order to

extract valuable information on pedestrian objects from image

data [1]-[5]. The first category of approaches require two

different procedures: detecting parts of a pedestrian object and

combining them for detecting an entire pedestrian object[1].

The second category of approaches first require finding low

level features within a target window and then determining if

the target window contains a pedestrian object by using some

statistical characteristics of the features [2].

A pedestrian object detection method proposed in this paper

is based on CNN(Centroid Neural Network) and SIFT (Scale

Invariant Feature Transform) [4] features and this method can

be considered as one of the second category methods. The

SIFT introduced by Lowe is invariant to scale, orientation, and

view point. The SIFT has been widely accepted since it

performs well especially in image matching, stereo matching

and motion tracking[1]-[5].

For classifier design procedure, Centroid Neural Network

(CNN) is adopted in this paper [6]-[8]. When compared with

SVM(Support Vector Machine), CNN itself is an unsupervised

clustering algorithm with a stable and fast clustering feature.

Thao Nguyen, Kheon-Hee Lee, Chang-Sun Kim and Dong-Chul Park are

with the Department of Electronics, Myongji University, YongIn, Rep. of

KOREA (phone: +82-31-330-6756, fax: +82-31-3306977, parkd@mju.ac.kr) Soo-Young Min is with Software Device Research Center at Korea

Electronics Technology Institute, SongNam, Rep. of KOREA (e-mail:

minsy@keti.re.kr )

The organization of this short paper is constructed as follows:

Section II introduces a keypoint extraction method by using

SIFT and a review on CNN. Feature extraction with keypoints

is proposed in Section III. Experiments on INRIA Person

dataset are given in Section IV. Section V concludes this

paper.

II. FEATURE EXTRACTION AND CLASSIFICATION

A. Scale Invariant Feature Transform

Since SIFT can provide features invariant to scale, rotation,

illumination and viewpoint, it has been widely used for

obtaining important invariant features from image[1]-[5]. By

using the features by SIFT, an object matching operation

images can be achieved[5]. The following procedures are

required when SIFT is adopted:

1)Detection of Scale-space extrema: searches over all scales

and image locations: DoG(Difference of Gaussian method)

2)Localization of Keypoints: finds a detailed model to

determine location and scale and finds stable one by passing

through a contrast and edge test.

3)Assignment of Orientation: finds dominant orientations for

keypoint in order to archive the rotation invariant.

4) Creation of Keypoint descriptor: finds a descriptor based

on the histogram of gradient to represent each keypoint. Finally,

the descriptor is used for alleviating illumination changes.

B. CNN(Centroid Neural Network)

The CNN algorithm[6] is an unsupervised competitive

learning algorithm based on the classical k-means clustering . It

finds the centroids of clusters at each presentation of the data

vector. The CNN first introduces definitions of the winner

neuron and the loser neuron. When a data x is given to the

network at the epoch (k), the winner neuron at the epoch (k) is

the neuron with the minimum distance to x. The loser neuron at

the epoch (k) to x is the neuron that was the winner of x at the

epoch (k-1) but is not the winner of x at the epoch (k). The CNN

updates its weights only when the status of the output neuron

for the presenting data has changed when compared to the

status from the previous epoch.

When an input vector x is presented to the network at

iteration n, the weight update equations for winner neuron j and

loser neuron i in CNN can be summarized as

Pedestrian Object Detection by Using

Centroid Neural Network

Thao Nguyen, Kheon-Hee Lee, Chang-Sun Kim, Dong-Chul Park, and Soo-Young Min

T

International Journal of Computer Science and Electronics Engineering (IJCSEE) Volume 2, Issue 2 (2014) ISSN 2320–4028 (Online)

80

(a) (b) Fig. 1 Distribution of keypoints: (a) Negative keypoints (b)Positive

key points

Fig. 2 Histogram feature components

( ) ( )

[ ( ) ( )] ( )

( ) ( )

( ) ( ) (2)

where ( ) ( ) are the winner and loser neurons with

and data, respectively.

The CNN has several advantages over conventional

clustering algorithms such as SOM or k-means algorithm when

used for clustering and unsupervised competitive learning. The

CNN requires neither a predetermined schedule for learning

gain nor the total number of iterations for clustering. It always

converges to sub-optimal solutions while conventional

algorithms such as SOM may give unstable results depending

on the initial learning gains and the total number of iterations.

More detailed description on the CNN can be found in [6]-[8].

III. KEYPOINTS FEATURE

With extracted keypoints and locations of keypoints for

backgrounds (negative images) and pedestrian objects (positive

images) as shown in Fig. 1 and Fig.2, respectively, we found

that there exists distinctive difference between distributions of

keypoints of positive images and those of negative images.

The keypoints of positive images mostly spread out more

widely in the image space than those of negative image. Based

on this observation, we are able to quantify this difference with

the feature descriptor as shown in Fig. 3. The extracted feature

is the orientation histograms for accumulated magnitudes of

Euclidean distance from each keypoint to the center point at the

relative angle between each keypoint and the center point. In

experiments, the best number of histogram bins is found to be

12. In our experiments, the numbers of keypoints for negative

and positive images are set to 112 and 158, respectively.

Fig. 3 Positive training data images[10]

Fig. 4 Negative training data images[10]

IV. EXPERIMENTS AND RESULTS

INRIA Person dataset is used for experiments[10]. 1,500

images for positive data and 1,500 images for negative data are

obtained. SIFT is applied to each of the fixed set of 20,000

patches randomly sampled from images. Then these SIFT

keypoint locations are processed to obtain the histogram of

orientation for keypoints feature. These histogram features for

each positive and negative data are passed through CNN for

clustering process. CNN is set to produce two clusters which

are corresponding to negative and positive.

In this procedure, when we utilize the fact that the number of

keypoints for positive imags is much larger than those for

negative image, we eliminate the windows which do not have

enough number of keypoints.

In order to evaluate the proposed method, the detection

accuracy and training speed for the proposed CNN method are

compared with the conventional SVM method The results are

summarized in Table I. As can be seen from Table I, detection

accuracy of CNN on the testing dataset is about 90.12% while

SVM shows almost perfect 99.96% accuracy. These results

come from the fact that CNN is a unsupervised learning

algorithm while SVM is a supervised learning algorithm. On

the other hand, however, CNN has a very important advantage

in training time over SVM: CNN finishes its training almost

instantly while SVM requires more than 3.28 hours for the

same training data on our PC environment (Intel Core Quad

2.33GHz, 4GB of RAM).

TABLE I

DETECTION ACCURACY AND TRAINING SPEED

CNN SVM

Accuracy 90.12% 99.96%

Training Time 0.92 s 3.28 h

International Journal of Computer Science and Electronics Engineering (IJCSEE) Volume 2, Issue 2 (2014) ISSN 2320–4028 (Online)

81

Fig. 5 Experimental results on INRIA dataset

V. CONCLUSIONS

In this paper, we propose a classifier model for pedestrian

object detection using CNN. The method combines SIFT and

CNN to produce an automatic pedestrian detection system. The

proposed method utilizes the observation of keyponts

distributions for positive data and negative data. Based on the

observation, we formulate a histogram feature for the

orientation and magnitude of feature points. The proposed

method is evaluated with INRIA data set. The results show that

the proposed method can detect pedestrian objects with

acceptable detection accuracy. The proposed method has the

advantageous features of both SIFT and CNN: extremely fast

training time and scale invariant feature. The proposed method

is acceptable for system that requires real time performance or

needs to update the training database quickly.

ACKNOWLEDGMENT

This work was supported by the IT R\&D program of The

MKE/KEIT (10040191, The development of Automotive

Synchronous Ethernet combined IVN/OVN and Safety control

system for 1Gbps class).

REFERENCES

[1] M. Brown and D. G. Lowe, “Recognising panoramas,” in Proc. IEEE Int.

Conf. Computer Vision, 2003, vol. 2, pp. 1218-1225. [2] M. Brown and D. G. Lowe, “Invariant Features from Interest Point

Groups,” Dep. of Computer Science, University of British Columbia,

Vancouver, Canada. [3] A. Tetal, “Face Description with Local Binary Patterns,” IEEE Trans.

Pattern Analysis and Machine Intelligence, vol.28, no.12, pp.2037-2041,

2006. [4] D. G.Lowe, “Distinctive Image Features from Scale-Invariant,”

International Journal of Computer Vision, vol. 60, no. 2 , pp. 91-110,

2004. [5] H. Zhou, Y. Yuan, and C. Shi, “Object tracking using SIFT features and

mean shift,” Computer Vision and Image Understanding, vo. 113, no. 3, pp. 345-352, March 2009.

[6] Dong-Chul Park, “Centroid Neural Network for Unsupervised

Competitive Learning,” IEEE Trans. Neural Networks, vol. 11, no. 2, pp. 520-528, May, 2000.

[7] Dong-Chul Park and Young-Jun Woo, “Weighted centroid neural

network for edge reserving image compression,” IEEE Trans. Neural Networks, vol. 12, no. 5, pp.1134-1146, March 2001.

[8] Dong-Chul Park, Oh-Hyun Kwon, and Jio Chung, “ Centroid neural

network with a divergence measure for gpdf data clustering,” IEEE Trans. Neural Networks, vo. 19, no. 6, pp. 948-957, June 2008.

[9] C. Cortes and V. Vapnik, “Support-Vector Network,” Machine

Learning, vo. 20, pp. 273-297, 1995. [10] http://pascal.inrialpes.fr/data/human/

Thao Nguyen received the B.S. degree in Computer Engineering from Ho Chi Minh City University of Technology in 2010 and the M.S. degree in

Electronics Engineering from MyongJi University. His research interests

include pattern recognition, deep learning, neural networks, and object recognition

Kheon-Hee Lee received the B.S. degree in Electronics Engineering from

MyongJi University, Korea, in 2013. He is pursuing his M.S. degree in Electronics Engineering at Intelligent Computing Research Lab at MyongJi

University. His research interests include pattern recognition, deep learning, neural networks, and object recognition

Dong-Chul Park (M’90-SM’99) received the B.S. degree in electronics

engineering from Sogang University, Seoul, Korea, in 1980, the M.S. degree in electrical and electronics engineering from the Korea Advanced Institute of

Science and Technology, Seoul, Korea, in 1982, and the Ph.D. degree in

electrical engineering, with a dissertation on system identifications using artificial neural networks, from the University of Washington (UW), Seattle, in

1990. From 1990 to 1994, he was with the Department of Electrical and

Computer Engineering, Florida International University, The State University of Florida, Miami. Since 1994, he has been with the Department of Electronics

Engineering, MyongJi University, Korea, where he is a Professor. From 2000

to 2001, he was a Visiting Professor at UW. He is a pioneer in the area of electrical load forecasting using artificial neural networks. He has published

more than 130 papers, including 40 archival journals in the area of neural

network algorithms and their applications to various engineering problems including financial engineering, image compression, speech recognition,

time-series prediction, and pattern recognition. Dr. Park was a member of the

Editorial Board for the IEEE TRANSACTIONS ON NEURAL NETWORKS from 2000 to 2002.

Soo-Young Min received the B.S. degree in Electronics Engineering from

Inha University, Incheon, Korea, in 1987. Since 1993, he has been with Software Device Research Center at Korea Electronics Technology Institute,

Song Nam, Korea, where he is a senior researcher. He has been involved in

numerous research projects on wireless communication protocol and system software. His research interests include vehicular communication network,

wireless communication protocol, and system software design.

International Journal of Computer Science and Electronics Engineering (IJCSEE) Volume 2, Issue 2 (2014) ISSN 2320–4028 (Online)

82